HP reportedly claims its now-orderable ZGX Fury combines a GB300 Superchip and 748GB of unified memory to support shared departmental or edge inference without a data center, expanding turnkey capacity for large local models.
state: corroboratedheat: lowuncertainty: mediumconvergesscott: mediumlocal-inference ai-infrastructure ai-workstationsHPNVIDIARed Hat
What is this?
HP's ZGX Fury is a rack-mountable 'AI station' — effectively NVIDIA's DGX Station platform in HP trim — now orderable, built on the GB300 Grace Blackwell Ultra Desktop Superchip with 748GB of coherent unified memory (496GB LPDDR5X on the CPU at ~396–400GB/s plus 252GB HBM3e on the GPU at 7.1TB/s) and up to 20 petaFLOPS FP4, marketed for shared departmental/edge inference of models up to ~1T parameters (FP4-quantized, per HP's own footnote) and fine-tuning of 100B-class models without data-center infrastructure, with a planned but not-yet-shipping Red Hat AI Factory integration. The orderability and specs are corroborated by HP's own pages, StorageReview's report, and an HN thread; the practical inference-performance and cost claims remain vendor-side, and the ~$100k-class price skepticism in HN discussion is unresolved. The Asus ExpertCenter Pro ET900N G3 — a GB300 station with an independent hands-on review — makes desktop-GB300 a multi-vendor category rather than an HP-only SKU.
Why it matters to Scott
HP/NVIDIA/Asus have independently productized the shared departmental local-inference pattern Scott already runs on gamepc (the shared Ollama endpoint at :11434) and prices in his platform-economics/TCO positions — a dated receipt that the 'inference without a data center' pattern now has vendor-scale turnkey form, confirmed as a multi-vendor category by the Asus hands-on rather than an HP-only claim. The unresolved questions land exactly on his territory: whether the 252GB HBM3e / 496GB LPDDR5X split actually permits practical FP4 trillion-parameter serving (his hardware-aware-inference usable-VRAM question), and whether ~$100k-class pricing beats hosted tokens (his GPU-server-vs-token-fees comparison), which also gives the AI consulting practice a concrete departmental-appliance option to advise on — pending StorageReview's independent review, this stays medium rather than high.
dev:project.gamepcdev:concept.hardware-aware-local-inferenceip:concept.platform-economicswork:concept.ai-consulting-practiceradar:concept.ai-workstationsradar:concept.gpu-memoryradar:concept.inference-economicsradar:apple-m5-ultra-local-inferenceradar:amd-threadripper-halo-stationradar:ssd-llama-trillion-parameter-local-inference
queries asked of Scott's wikis
- shared local inference endpoint for team/department — GPU server economics vs token fees
- unified memory vs HBM split — usable VRAM for very large local model serving
- local model hosting hardware: Mac Studio / multi-GPU workstation / appliance comparisons
- on-prem vs cloud inference TCO and when self-hosting wins
- edge/departmental AI appliance — inference without data center
- FP4 quantization serving trillion-parameter models locally
Measured heat
now 0 pts/hpeak 0 pts/hcomments 0/hpeers p14momentum: steady2 platformsage 794h
points/hour across evidence · reading as of 2026-10-12 02:59:37.977291+11:00 · deterministic, not a model opinion
How the heat travelled
pace: p67 vs 519 stories at the 720h mark (now 794h old) — ahead of applied-compute-training-serving-platform (1.0x), behind openai-german-wiki-incident (1.0x)
Evidence (3) — ⭐ canonical anchor
| source | object | author | score | comments |
|---|
| 🟧 hn | HP ZGX Fury Is Now Orderable: GB300 Superchip, 748GB Unified MemoryRetrieved article excerptOpen article · Retrieved 2026-09-14T11:22:04.442730+00:00 # HP ZGX Fury Is Now Orderable: GB300 Superchip, 748GB Unified Memory, and a Red Hat AI Factory Plan for the Edge
by Brian Beeler
on September 9, 2026
[Consumer](https://www.storagereview.com/consumer) ◇
[Workstation](https://www.storagereview.com/consumer/workstation)
HP’s ZGX Fury AI station is now available to order, and HP paired the availability news with a collaboration with Red Hat and NVIDIA to put Red Hat AI Factory with NVIDIA on top of it. The ZGX Fury is HP’s take on NVIDIA’s DGX Station design, built around the GB300 Grace Blackwell Ultra Desktop Superchip with 748GB of unified memory and up to 20 petaFLOPS of FP4 compute, and HP is positioning it less as a personal workstation than as a shared inference box that a department, a factory floor, or a branch office can run without a data center behind it. We have one in the lab now, so a full review is coming; this is what HP has said so far.
HP ZGX Fury AI station tower in the StorageReview lab, showing the front mesh panel, ZGX badge, and front USB and audio ports
**Engage with StorageReview**
[Newsletter](https://www.storagereview.com/storage_newsletter) | [YouTube](https://www.youtube.com/user/StorageReview "Opens in a new window") | Podcast [iTunes](https://podcasts.apple.com/gb/podcast/storagereview-com-storage-reviews/id1060681115 "Opens in a new window")/[Spotify](https://open.spotify.com/show/1y6VnznABhHeOSMOmbDTz0 "Opens in a new window") | [Instagram](https://www.instagram.com/storagereview/ "Opens in a new window") | [Twitter](https://twitter.com/storagereview "Opens in a new window") | [TikTok](https://www.tiktok.com/@storagereview "Opens in a new window") | [RSS Feed](https://www.storagereview.com/rss.xml)
Previous post: [Samsung High NA EUV DRAM Set for 2028, With 12-Inch Photomasks and a Mistral AI Series D Lead](https://www.storagereview.com/news/samsung-high-na-euv-dram-set-for-2028-with-12-inch-photomasks-and-a-mistral-ai-series-d-lead)
Next post: [Scality Maestro 1.1 Manages ARTESCA Fleets From One Console With No VPN Into Customer Sites](https://www.storagereview.com/news/scality-maestro-1-1-manages-artesca-fleets-from-one-console-with-no-vpn-into-customer-sites) | rbanffy | 41 | 54 |
| 🟧 echo.other ⭐ | HP’s press release states that “HP ZGX Fury is now available to order” and announces HP’s collaboration with Red Hat and NVIDIA on an enterp | HP Inc. | — | — |
| 🟧 hn | Asus ExpertCenter Pro ET900N G3 Review: The GB300 DGX Station Gets Handles | rbanffy | 5 | 0 |
Interpretation history
2026-09-24T14:01:16Z
grounded: converges/medium — HP/NVIDIA/Asus have independently productized the shared departmental local-inference pattern Scott already runs on gamepc (the shared Ollama endpoint at :11434
2026-09-24T13:55:00Z
Asus shipping an equivalent GB300 station (ExpertCenter Pro ET900N G3) with an independent hands-on review makes the turnkey desktop-GB300 inference box a real multi-vendor category rather than a one-off HP claim — corroborated at category level. HP-specific performance and cost remain vendor-only claims; the case's meaning shifts from 'HP's new SKU' to 'an emerging desktop-GB300 category Scott's local-inference practice can evaluate'.
2026-09-24T13:28:10Z
evidence attached: hn.story.49829803 — Independent hands-on review of a GB300 workstation confirms the same turnkey desktop-GB300 category episode the HP ZGX Fury case tracks, now spanning vendors.
2026-09-14T11:26:05Z
grounded: converges/medium — HP’s turnkey shared-inference station converges with Scott’s gamepc shared local-model endpoint and adds a higher-memory hardware option to evaluate against his
2026-09-14T11:23:18Z
origin walked (codex/luna, conf 0.99): anchor hn.story.49694905 -> echo.other.212e8e91f4 by HP Inc.
2026-09-14T11:22:14Z
case created — Reported orderability and a unit already in StorageReview's lab make this a concrete hardware episode, with practical inference performance still unvalidated.
Decision trace
- 10-07 18:17review_dormantscheduled targets exhausted or 28 quiet days
- 10-07 18:17drop_targetsquiet through full ladder or over cap 8
- 09-25 00:01repriceAsus shipping an equivalent GB300 station (ExpertCenter Pro ET900N G3) with an independent hands-on review makes the turnkey desktop-GB300 inference box a real multi-vendor category rather than a one-
- 09-25 00:01groundHP/NVIDIA/Asus have independently productized the shared departmental local-inference pattern Scott already runs on gamepc (the shared Ollama endpoint at :11434) and prices in his platform-economics/T
- 09-24 23:55review_reactivatedAsus shipping an equivalent GB300 station (ExpertCenter Pro ET900N G3) with an independent hands-on review makes the turnkey desktop-GB300 inference box a real multi-vendor category rather than a one-
- 09-24 23:28attachIndependent hands-on review of a GB300 workstation confirms the same turnkey desktop-GB300 category episode the HP ZGX Fury case tracks, now spanning vendors.
- 09-24 23:27propose_attachIndependent hands-on review of a GB300 workstation confirms the same turnkey desktop-GB300 category episode the HP ZGX Fury case tracks, now spanning vendors.
- 09-24 16:43review_dormantscheduled targets exhausted or 28 quiet days
- 09-24 16:43drop_targetsquiet through full ladder or over cap 8
- 09-15 04:45review_screenThe change adds a joking comment and removes or lightly edits existing price and memory-capacity opinions without introducing new facts or evidence.
- 09-14 22:37review_screenThe new comments mainly repeat cost concerns and offer unverified opinions about usable memory and workload suitability; they add no credible implementation result or consequential new fact.
- 09-14 22:20sensor_dirtycomment_update
- 09-14 21:26groundHP’s turnkey shared-inference station converges with Scott’s gamepc shared local-model endpoint and adds a higher-memory hardware option to evaluate against his hardware-aware local-inference practice
- 09-14 21:23promote_anchororigin walk conf 0.99
- 09-14 21:22createReported orderability and a unit already in StorageReview's lab make this a concrete hardware episode, with practical inference performance still unvalidated.